Skip to content

Security: ProReka/multica-runtime

Security

SECURITY.md

Security Policy

Supported versions

Security fixes land on main and in the next tagged release. Pin a release tag or digest in production; :latest receives fixes first but moves without notice.

Version Supported
latest release ✅
main builds ✅
older tags best effort

Reporting a vulnerability

Do not open a public issue.

Use GitHub's private vulnerability reporting:

https://github.com/ProReka/multica-runtime/security/advisories/new

Include:

  • affected image and tag/digest;
  • what an attacker can do (read a secret, escape the container, …);
  • reproduction steps, ideally a minimal Dockerfile or command;
  • whether the issue is in this project or in an upstream component (Claude Code, OpenCode, Multica CLI, Debian package, …).

What to expect

  • Acknowledgement within 3 business days.
  • An assessment and a remediation plan within 10 business days.
  • Credit in the advisory unless you ask otherwise.

If the issue is in an upstream tool, we will still help coordinate, but the fix may need to land upstream first. We can pin or disable a component in the image while waiting.

Scope

In scope:

  • secrets leaking into an image layer, logs, or the repository;
  • privilege escalation or container escape caused by this project's configuration;
  • an image running as root, or requiring a capability it should not;
  • supply-chain issues in the build (unpinned download, missing checksum);
  • the autoscaler scaling the wrong service or bypassing its guards.

Out of scope:

  • prompt injection through task content — the agent executes code by design;
  • a compromised AI provider or Multica account;
  • vulnerabilities in a task's own code;
  • the Docker socket mount in the optional Swarm autoscaler (documented as root-equivalent; use the ECS adapter if that is not acceptable).

Hardening checklist for operators

  • Pin images by digest.
  • Scope MULTICA_TOKEN and GH_TOKEN to what the agent needs.
  • Restrict egress if your environment allows it.
  • Do not mount host directories that contain unrelated credentials.
  • Do not mount the Docker socket unless you run the Swarm autoscaler and trust every task on that host.

There aren't any published security advisories